Introduction

Europe is currently an epicentre of the COVID-19 pandemic.[1] In response to widespread community spread and to urgently reduce infection rates of the COVID-19 disease in the population, many European governments have imposed strict “lockdown” measures. These lockdown measures are unprecedented and compel (or force) most of the population to stay in their homes, work from home if possible, and businesses such as restaurants, shopping malls, schools, universities, playgrounds, etc. are closed. Essential services remain open and generally include commercial activities around food, medication, and utilities.

The Swiss Federal Council announced a country-wide lockdown on March 16, 2020 which would begin on March 17 and provisionally end on April 20 (but has subsequently been extended to April 26).[2] The Swiss lockdown is somewhat more lenient than those seen in other European countries. Due to Switzerland’s federalist structure, there are also some differences among the different cantons. For example, the southern canton of Ticino boarding Italy introduced many measures earlier than the rest of the country. This was because the area experienced an earlier increase in infections due to the proximity to the northern Italy case-cluster.

The lockdowns experienced in Europe have effectively acted as a switch where economic activity has all but stopped within a very short amount of time. The financial and economic consequences of such measures will likely be dire. A small positive which comes from such actions is that air quality should improve rather quickly due to emissions of atmospheric pollutants being dramatically reduced across almost all of Europe. There have been many news articles showing this effect, and the satellite imagery over China and Europe have been particularly striking.[3–6] Here, we will explore the situation in Switzerland with ambient monitoring data from the National Air Pollution Monitoring Network (NABEL).

The importance of weather

A discussion which has been lacking in most of the news articles, and is very important to consider, is the influence of weather when investigating changes in air pollutant concentrations. It is obvious that emissions of pollutants would have reduced throughout Europe due the the COVID-19 lockdown measures from curtailed economic activities. This would have resulted in lower concentrations and better air quality in many locations. However, the same effect of decreasing concentrations would be observed if, for example, wind speed and atmospheric dispersion had increased. This is a classic issue and question in air quality interpretation: “Are the changes observed due to emissions or weather?” Here, we take into account changes in weather so we can more robustly determine what effect the lockdown had on air quality. A similar analysis has been conducted in the United Kingdom by David Carslaw, a previous supervisor and a colleague.[8]

All data and analyses are preliminary at this stage and research is ongoing.

Analysis framework

Hourly nitrogen dioxide (NO2), oxides of nitrogen (NOx), and ozone (O3) data from many NABEL ambient monitoring sites across Switzerland were used.[9] Basic details of the sites can be found in Table 1 and the locations of the sites are shown in Figure 1. Not all NABEL sites were included in the analysis. Sites were selected based on what pollutants are monitored and their site type (classification of what type of environment the sites are located in, for example, roadside or rural). The sites’ observations were used to train random forest machine learning models[10] to predict pollutant concentrations based on meteorological/weather measurements. The random forest models also used time variables as inputs because pollutants vary over different time scales, for example, by season and throughout the day. Complete and thorough details of the methods can be found elsewhere.[11, 12] The approach has been used in Switzerland before, but for a different application.[13]

Table 1: Information about the NABEL Swiss air quality monitoring sites.
Site Site (NABEL code) Site name Lat. Long. Elevation (m) Site type Site area
ch0001g JUN Jungfraujoch 46.547 7.985 3578 Special high alpine Rural
ch0002r PAY Payerne 46.813 6.944 489 Background Rural
ch0003r TAE Tänikon 47.480 8.905 538 Background Rural
ch0004r CHA Chaumont 47.050 6.979 1136 Background Rural
ch0005a DUE Dübendorf-Empa 47.403 8.613 432 Background Suburban
ch0005r RIG Rigi-Seebodenalp 47.067 8.463 1031 Background Rural
ch0008a BAS Basel-Binningen 47.541 7.583 316 Background Suburban
ch0010a ZUE Zürich-Kaserne 47.378 8.530 409 Background Urban
ch0011a LUG Lugano-Università 46.011 8.957 280 Background Urban
ch0028a LAU Lausanne-César-Roux 46.522 6.640 530 Traffic Urban
ch0031a BER Bern-Bollwerk 46.951 7.441 536 Traffic Urban
ch0032a HAE Härkingen-A1 47.312 7.821 431 Traffic Rural
ch0033a MAG Magadino-Cadenazzo 46.160 8.934 203 Background Rural
ch0052a SIO Sion-Aéroport-A9 46.220 7.342 483 Traffic Rural
ch2000e BER Beromünster 47.190 8.175 797 Background Rural
ch2001e DAV Davos-Seehornwald 46.802 9.839 1637 Background Rural

Figure 1: Locations of Switzerland’s NABEL air quality monitoring sites.

Data between January 1 2019 and February 29 2020 were used for training of the random forest models. Data from March 1 2020 onwards were not used for model training, however, the meteorological data were used to predict pollutant concentrations. These predictions after March 1 can be thought of as a “business as usual” scenario and gives an estimate on what pollutant concentrations would have been without the decrease in emissions resulting from reduced economic activity due to the lockdown. Predictions can be compared with the observed values to see how the measurements differ. The models used are regression models so they generalise and are unable to capture minima and maxima of concentrations of some pollutants particularly well. However, if we look at errors over time, and if the business as usual scenario is true, the errors will bounce around zero. In contrast, if the system changes, the errors will diverge from zero. This is the basis of the analysis.

Results and discussion

First checks

Time series of pollutant concentrations at selected monitoring sites in March 2020 are shown in Figure 2. It seems that concentrations of NO2, NOx, and O3 did not drastically change after the lockdown was implemented. There are however examples, such as NO2 and NOx at Lugano-Università, where concentrations have visibly decreased after the lockdown, indicating that concentrations and lockdown measures are connected. Therefore, additional, more in-depth exploration is needed to help explain the patterns seen in Figure 2.

Time series of daily NO~2~, NO~x~, and O~3~ in March 2020 for selected of monitoring sites in Switzerland.

Figure 2: Time series of daily NO2, NOx, and O3 in March 2020 for selected of monitoring sites in Switzerland.

When starting an analysis such as this, it is important to consider what the general state-of-play of weather conditions have been for the analysis period. The weather in the first three months of 2020 (winter and the first month of spring) has been mild in most of Switzerland.

In Figure 3, it can been seen that average air temperatures in the beginning of 2020 have been higher in almost all locations plotted compared to the last few years. In some locations, such as Basel and Zürich, wind speed has also been unusually high. The warmer temperatures and higher wind speeds have likely led to a less stable planetary boundary layer (the lowest portion of the atmosphere). Although this may not have been universal across all of Switzerland, the dispersion characteristics of the atmosphere have likely been enhanced in the January–March 2020 period compared to previous years in many locations. These observations suggest that emissions of pollutants may have been lower than the last previous years, and in some locations there would have been enhanced dispersion.

Means of meteorological variables for a number of locations across Switzerland between January, and March for 2017, 2018, 2019, and 2020.

Figure 3: Means of meteorological variables for a number of locations across Switzerland between January, and March for 2017, 2018, 2019, and 2020.

Mean pollutant concentrations for the same periods (between January and March for 2017, 2018, 2019, and 2020) are displayed in Figure 4. Here, we can see for NO2 and NOx, that concentrations are generally lower in 2020 than in previous years. In some locations such as Bern-Bollwerk, Härkingen-A1, and Payerne the mean NO2 and NOx concentrations in 2020 are much lower than in previous years. However, for O3, this is not the case.

Pollutant concentration means for selected monitoring sites in Switzerland between January, and March for 2017, 2018, 2019, and 2020.

Figure 4: Pollutant concentration means for selected monitoring sites in Switzerland between January, and March for 2017, 2018, 2019, and 2020.

The aggregations in 2020 shown in Figure 4 include two weeks of lockdown measures, so this question can be asked: "Are the lockdown measures responsible for the observed decreases?’’ Based on the meteorological data shown in Figure 3, it is expected that concentrations may have been lower due to a mild winter and increased atmospheric dispersion in many locations in Switzerland. Therefore, further work is required to disentangle the influence of the lockdown measures and the potentially confounding meteorological factors.

Modelling pollutant concentrations for March 2020

After the random forest models were trained, checked for adequate skill, and used to predict concentrations of NO2, NOx, and O3 for March 2020, the observed concentrations were compared to those which were predicted. Figure 5 shows daily time series of observed and predicted concentrations for March 2020 for the sites included in the analysis. For some sites, such as Bern-Bollwerk and Lugano-Università, a clear divergence is seen between the NO2 observed and predicted values after the March 16 lockdown date. The observed concentrations were lower than the predicted values suggesting that actual NO2 concentrations were lower than those based on a business as usual scenario. The differences (delta) between the observed and predicted values are also shown in Figure 6.

Observed and predicted NO~2~, NO~x~, and O~3~ concentrations for March 2020 for selected Switzerland's monitoring sites.

Figure 5: Observed and predicted NO2, NOx, and O3 concentrations for March 2020 for selected Switzerland’s monitoring sites.

Delta of observed and predicted NO~2~, NO~x~, and O~3~ concentrations for March 2020 for the selected Switzerland's monitoring sites

Figure 6: Delta of observed and predicted NO2, NOx, and O3 concentrations for March 2020 for the selected Switzerland’s monitoring sites

To more clearly expose the differences between the observed and predicted values, cumulative sums (cumsum) were calculated for the selected sites’ pollutants. This technique simply aggregates the deltas/differences over time and allows for identification of both the direction and time when the divergence occurred. Cumulative sums of NO2, NOx, and O3 for the selected sites are displayed in Figure 7.

Cumulative sum for observed and predicted NO~2~, NO~x~, and O~3~ values for March 2020 for selected Switzerland's monitoring sites.

Figure 7: Cumulative sum for observed and predicted NO2, NOx, and O3 values for March 2020 for selected Switzerland’s monitoring sites.

Figure 7 is a key figure in many ways because the changes shown are relative to the business as usual scenario. First, it shows that NO2 and NOx have decreased across most of Switzerland, but for the rural sites (Beromünster and Payerne), this decrease has been much less dramatic. For Bern-Bollwerk, Lausanne-César-Roux, and Härkingen-A1 (all traffic sites), the divergence from the business as usual scenario occurs immediately at the start of March which suggests that concentrations of NO2 and NOx were lower than what the model predicts two weeks before the lockdown was announced. This could be explained by NOx emissions already being lower than normal two weeks before the lockdown, or limited skill of the model from using the surface meteorological variables for training—or perhaps a combination of both. These features in Figure 7 link back to the interpretation of the simple aggregations shown in Figure 3 and Figure 4. Because these three sites are classed as traffic sites (Table 1), the decreases can be interpreted as lower emissions of these pollutants in close proximity to the monitoring sites.

Figure 7 also suggests that O3 concentrations have been higher in March 2020 than the business as usual scenario. O3 is not a primary (directly emitted) pollutant and is generated in the atmosphere. O3 and NOx are linked by a transformation cycle where a component of NOx is transformed to O3 in the presence of ultra-violet light. Because of this relationship, if NOx concentrations decrease, O3 concentrations increase which is what is observed in Figure 7. O3 concentrations have remained well below the legal limit for this pollutant across Switzerland for the analysis period.

Looking at the differences after March 16

For the period after March 16 2020, the differences between the observed and predicted values can be aggregated and transformed into percentage change. Table 2 shows these calculations and a plot of the observed and predicted means are displayed in Figure 8. Table 2 demonstrates that NO2 and NOx concentrations have decreased by up to 46 and 57 % respectively (at Lugano-Università) based on the scenario modelling technique. Care is needed with interpreting these results however because the uncertainty calculations have not been conducted yet. Additionally, Figure 7 demonstrates that some models showed immediate departure from the observed time series when used to predict at the start of March 2020. This may led to an overestimation of the changes seen in Table 2. Beromünster is the only site with estimated increases in NO2 and NOx. The differences between the absolute values are however small: 1.3 and 1.6 µg m-3 for NO2 and NOx respectively (Table 2; Figure 8). These differences are likely to be in the range of uncertainty of the modelling estimates, but further investigation is required to fully to explain this behaviour.

Inversely, O3 concentrations have increased in almost all locations across Switzerland in response to lower NOx which is to be expected due to the chemical relationship between NOx and O3 within the troposphere (lower atmosphere). A figure visualising the percentage changes is available in Figure 9. Please note that the percentage change calculations are relative to one another and need to be interpreted alongside the concentrations (these data are in Table 2 and Figure 8).

Table 2: Observed and predicted concentration means, deltas, and percentage change for the period after March 16 2020 for Switzerland’s selected monitoring sites.
Site name Variable Observed Predicted Observed-predicted delta Predicted-observed percentage change
Lugano-Università NO2 16.7 31.3 -14.5 -46.4
Bern-Bollwerk NO2 24.9 38.8 -13.9 -35.7
Lausanne-César-Roux NO2 22.5 34.2 -11.7 -34.1
Härkingen-A1 NO2 20.9 29.0 -8.0 -27.7
Magadino-Cadenazzo NO2 13.5 18.6 -5.1 -27.4
Zürich-Kaserne NO2 23.4 28.4 -5.0 -17.7
Basel-Binningen NO2 16.4 19.1 -2.7 -14.2
Dübendorf-Empa NO2 23.7 25.9 -2.2 -8.5
Payerne NO2 11.1 11.9 -0.8 -7.1
Beromünster NO2 8.7 7.4 1.3 17.2
Lugano-Università NOx 18.6 43.4 -24.8 -57.1
Bern-Bollwerk NOx 41.2 75.1 -33.9 -45.2
Lausanne-César-Roux NOx 32.1 57.4 -25.3 -44.0
Magadino-Cadenazzo NOx 16.3 26.4 -10.1 -38.3
Härkingen-A1 NOx 39.1 56.0 -17.0 -30.3
Zürich-Kaserne NOx 28.8 38.8 -10.0 -25.7
Dübendorf-Empa NOx 32.9 43.0 -10.1 -23.6
Basel-Binningen NOx 19.9 24.1 -4.2 -17.3
Payerne NOx 12.7 15.1 -2.4 -16.1
Beromünster NOx 9.5 8.0 1.6 19.6
Dübendorf-Empa O3 50.8 55.9 -5.1 -9.1
Zürich-Kaserne O3 54.6 57.0 -2.4 -4.2
Payerne O3 61.5 63.4 -1.9 -3.0
Basel-Binningen O3 59.8 59.7 0.1 0.2
Härkingen-A1 O3 52.8 52.4 0.4 0.8
Beromünster O3 76.8 75.4 1.4 1.9
Lausanne-César-Roux O3 57.5 53.7 3.9 7.2
Magadino-Cadenazzo O3 57.8 51.3 6.5 12.7
Bern-Bollwerk O3 50.2 44.1 6.1 13.8
Lugano-Università O3 61.8 49.8 12.0 24.1
Observed and predicted concentration means for the period after March 16 2020 for Switzerland's selected monitoring sites.

Figure 8: Observed and predicted concentration means for the period after March 16 2020 for Switzerland’s selected monitoring sites.

Predicted and observed percentage change for the period after March 16 2020 for Switzerland's selected monitoring sites. Sites are ordered by percentage change. Note that the percentage change is a relative change and care is needed in the interpretation.

Figure 9: Predicted and observed percentage change for the period after March 16 2020 for Switzerland’s selected monitoring sites. Sites are ordered by percentage change. Note that the percentage change is a relative change and care is needed in the interpretation.

Final notes

A classic question in air quality data analysis is if changes in pollutant concentrations are caused by reduction of emissions or changes in weather. To robustly investigate the effect of the COVID-19 lockdown measures in Switzerland on air quality, a scenario modelling approach was applied using machine learning models. This approach allows for an estimate of what pollutant concentrations would have been if lockdown measures were not applied in March 2020. These estimates can be compared to what was observed and the differences explored.

The results indicated that NO2 and NOx concentrations have decreased in most locations in Switzerland by up to 46 and 57 % respectively due to the lockdown measures. Care is needed with interpreting these results because of the relative nature of these calculations and the lack of uncertainty calculations at this stage. O3 on the other hand has generally increased across Switzerland due to the inverse relationship this pollutant has with NOx. These results outline the complexity of the atmospheric pollutant climate with different sources and the links among different species. This is something which has been stated in some news articles but has rarely been quantified.

Please note that the observations used and the data analysis are preliminary at this stage and research is on going. The results will be enhanced and revised regularly in the near future.

References

1. BBC. (2020). Coronavirus: Europe now epicentre of the pandemic, says WHO. Retrieved from https://www.bbc.com/news/world-europe-51876784

2. Federal Office of Public Health (FOPH). (2020). New coronavirus: Federal government measures. Retrieved from https://www.bag.admin.ch/bag/en/home/krankheiten/ausbrueche-epidemien-pandemien/aktuelle-ausbrueche-epidemien/novel-cov/massnahmen-des-bundes.html

3. Watts, J., & Kommenda, N. (2020). Coronavirus pandemic leading to huge drop in air pollution. Retrieved from https://bit.ly/2Rnkvbk

6. Ammann, K. (2020). Stickstoffdioxid und Ozon, Die Luft in der Schweiz ist sauberer—aber warum? Retrieved from https://www.srf.ch/news/schweiz/stickstoffdioxid-und-ozon-die-luft-in-der-schweiz-ist-sauberer-aber-warum

7. Wannan, O. (2020). Coronavirus: Traffic pollution plummets across the country during lockdown. Retrieved from https://www.stuff.co.nz/environment/climate-news/120763215/coronavirus-traffic-pollution-plummets-across-the-country-during-lockdown?cid=app-iPhone

8. Carslaw, D. (2020). Analysis of COVID-19 lockdown on UK local air pollution. Retrieved from https://ee.ricardo.com/news/analysis-of-covid-19-lockdown-on-uk-local-air-pollution

9. Federal Office for the Environment. (2017). National air pollution monitoring network (NABEL). Retrieved from https://www.bafu.admin.ch/bafu/en/home/topics/air/state/data/national-air-pollution-monitoring-network--nabel-.html

10. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. doi:10.1023/A:1010933404324

11. Grange, S. K., & Carslaw, D. C. (2019). Using meteorological normalisation to detect interventions in air quality time series. Science of the Total Environment, 653, 578–588. Retrieved from http://www.sciencedirect.com/science/article/pii/S004896971834244X

12. Grange, S. K. (2018). rmweather: Tools to Conduct Meteorological Normalisation on Air Quality Data. Retrieved from https://CRAN.R-project.org/package=rmweather

13. Grange, S. K., Carslaw, D. C., Lewis, A. C., Boleti, E., & Hueglin, C. (2018). Random forest meteorological normalisation models for Swiss PM\(_{10}\) trend analysis. Atmospheric Chemistry and Physics, 18(9), 6223–6239. doi:https://doi.org/10.5194/acp-18-6223-2018